PoloIQ turns raw water polo game film into structured, usable data — automatically. Point a camera at a game, and PoloIQ tracks every player, the ball, and the goal for the full match, identifies who's who by cap number, and converts that into player stats, positional heatmaps, and event logs — without a human spending hours breaking down film by hand.
Water polo has never had a real analytics product. Sports like basketball, soccer, and football have mature computer-vision tracking systems because they have decades of labeled video to train on. Water polo doesn't — players are submerged to the shoulders, everyone wears the same suit, cap numbers are tiny and often blurred, and no public dataset of labeled game footage exists. PoloIQ's core technical bet is solving that cold-start problem directly, which is what makes the rest of the product possible.
Who it's for
Recruit and evaluate players with objective, consistent data instead of gut feel — and break down opponent film automatically instead of by hand.
See individual positioning, movement patterns, and involvement trends over a season — the self-scouting data athletes in bigger sports already have.
Get consistent, structured data across every game of a season without hiring a dedicated video analyst.
What it does
- Automatic player & ball tracking across the full game, frame by frame.
- Cap-number identification — ties every tracked action to a specific player without manual tagging.
- Team & possession detection via jersey and cap color analysis.
- Heatmaps & positioning data, mapped to real pool coordinates rather than raw video pixels, so distances and positioning are physically meaningful.
- Pose & movement analysis — arm and body angles relevant to shooting and swimming technique.
- Shot placement mapping — where shots land relative to the goal, across a 6-zone grid (high / low × left / center / right).
- Automatic event detection — flags key moments like goals and turnovers so coaches can jump straight to the clip instead of scrubbing full-game footage.
How it works
PoloIQ is built as two connected systems: an annotator, which teaches the system what water polo looks like from raw, unlabeled footage, and a pipeline, which runs fast, trained models on new games.
Stage 1 — Annotator
Solves the cold-start problem: no existing labeled water polo dataset exists, so PoloIQ uses zero-shot foundation models (Grounding DINO for detection, SAM for segmentation) that can identify objects they were never explicitly trained on. It auto-labels raw footage across nine categories — player, ball, goal, cap, cap number, pool, referee/official, pool furniture, and line markers — deliberately including officials and pool furniture as their own categories so the system learns not to mistake a referee for a player, or a ladder for a goal.
Stage 2 — Pipeline
Those auto-generated labels train a fast, purpose-built detector (YOLOv8 with ByteTrack for tracking) that runs efficiently on new game footage. Downstream modules turn raw detections into meaningful stats:
Team ID via color clustering on caps and suits.
Cap-number reading to attribute actions to a specific player.
Homography mapping video pixels to real pool coordinates.
Body-mechanics model for shooting and swimming technique.
Goal-zone classifier for shot placement analysis.
This "auto-label with zero-shot models, then train a lightweight specialist model" approach is the core of the technology. It lets PoloIQ bootstrap accurate, sport-specific tracking with zero existing training data, and it keeps improving as more footage is processed — a strategy that could generalize to other underserved sports facing the same data gap.
Where things stand
PoloIQ is currently in pre-product R&D. The core computer-vision pipeline — detection, tracking, cap identification, pose, goal-zone mapping, and event detection — is fully built and has been validated end-to-end on real game footage, including a labeled corpus of over 9,000 frames from actual matches.
Testing against real footage has already surfaced and fixed several non-obvious failure modes — for example, confidently-wrong predictions from an older model incorrectly overriding correct-but-less-confident predictions from a newer one during label fusion. That's the kind of issue that only shows up once a system runs against real, messy game video rather than a clean benchmark.
The next phase: completing a full training pass across the expanded detection categories, and building the coach- and player-facing product interface on top of the existing pipeline.